feat: SIR correction for VIPosterior

[janfb]

Summary

VIPosterior.sample() had documented support for SIR debiasing via method="sir" and K parameters, but the implementation was never written. The sampling code was always just self.q.sample(sample_shape) - the method parameter was accepted but ignored.

The documentation was removed in June 2022 (12d59715), but tests continued passing these parameters because **kwargs silently swallowed them.

Question

Is SIR debiasing for VIPosterior something we want to implement (for real this time)?

Arguments for:

• Corrects approximation error in q without retraining
• Cheap post-hoc improvement
• Already implemented in ImportanceSamplingPosterior (could reuse logic)

Arguments against:

• Well-trained VI (especially fKL, IW methods) already minimizes this gap
• Users who need this can use ImportanceSamplingPosterior with trained q as proposal
• Was documented but never prioritized for implementation

any thoughts @manuelgloeckler @michaeldeistler ?

[satwiksps]

Hi @janfb @manuelgloeckler @michaeldeistler, I would like to work on this issue.

I believe while well-trained VI is powerful, SIR provides a cheap, post-hoc way to correct for approximation errors or mode-seeking behavior without the computational cost of retraining.

Proposed Implementation Plan:

1. Reusing Existing Logic: I will implement the SIR logic by utilizing the shared sampling_importance_resampling helper from sbi.samplers.importance.sir, ensuring consistency with ImportanceSamplingPosterior.
2. Robust Proposal Wrapper: Since VIPosterior flows can return log-probabilities with varied shapes (e.g., (N, 1)), I will implement a internal wrapper for the variational proposal to ensure log_prob is flattened before weight calculation. This prevents broadcasting errors/memory blowouts during the $log(p/q)$ step.
3. API Support: I will update sample() to explicitly handle the method="sir" argument and maintain backward compatibility for the legacy sample_with alias.
4. Verification: I will add a statistical regression test in tests/vi_test.py to verify that SIR sampling effectively reduces mean and variance errors on a known biased proposal.

Please let me know if this plan looks good to you, and I will open a PR!

[janfb]

Thank you @satwiksps for offering you help here!

I would wait for @manuelgloeckler 's call here as he implemented the original version.

My current take is that a how-to-guide or documentation on how to do the SIR correction using the ISPosterior could suffice.

Lets wait for Manuel's input - thanks! 🙏

[patelshivani2283-lab]

Hi everyone, thanks for raising this!

From a contributor’s perspective, I see value in SIR debiasing

A possible low-risk path forward could be:

• Implement SIR debiasing in VIPosterior.sample() by reusing or adapting the existing logic from ImportanceSamplingPosterior
• Keep it strictly opt-in (method="sir"), with clear documentation that it’s intended as a post-hoc correction rather than a replacement for better-trained VI
• Add a small test to ensure the parameters are no longer silently ignored

That said, I agree that users who need stronger guarantees can already use ImportanceSamplingPosterior, so I’d be happy to follow maintainers’ guidance on whether this should be implemented or explicitly rejected.

Looking forward to your thoughts!